Particle clustering in turbulence: Prediction of spatial and statistical properties with deep learning

Yan-Mong Chan, Natascha Manger, Yin Li, Chao-Chin Yang, Zhaohuan Zhu, Philip J. Armitage, Shirley Ho

Introduction

In the early stages of planet formation, the growth of micron-sized dust grains into planetesimals depends critically on their ability to cluster within turbulent gas. In protoplanetary disks, turbulence acts as a cosmic concentrator, driving particles into dense filaments and clumps. Traditionally, studying these dynamics requires Lagrangian simulations that track millions of individual particles interacting with a fluid grid. While highly accurate, these simulations are computationally intensive, limiting the exploration of diverse particle sizes and turbulence regimes. This research explores the use of deep learning—specifically 3D U-Nets—as a surrogate model to map gas behavior directly to particle clustering, offering significant potential for computational speedup.

Methodology

The researchers employed a 3D U-Net architecture, a convolutional neural network (CNN) designed for volumetric data. The model takes 3D grids of gas density and velocity as input to predict corresponding 3D grids of particle density and velocity.

A key innovation in this work is the training objective for particle motion. Rather than predicting absolute particle velocity (UNET-V), the authors developed a variant (UNET-R) that predicts relative velocity—the difference between particle and gas velocity. By focusing on the deviation from the fluid flow, the network can better capture the subtle physical interactions that drive clustering, rather than being dominated by the bulk motion of the gas.

Training data was generated using the Athena++ high-performance hydrodynamics code at a resolution of 256³ grid cells. The dataset comprised 252 simulation snapshots, augmented through random cropping, rotations, and parity flips to ensure the model learned underlying physical patterns rather than grid-specific features.

Spatial Fidelity and Visual Results

The U-Net models successfully captured the complex, filamentary structures characteristic of particle clustering in turbulence. Qualitatively, the predicted density fields closely mirror the ground truth, although the machine-learning outputs tend to be slightly smoother than the high-resolution simulations.

Comparison of particle density slices between the ground truth (Athena++) and the UNET-V and UNET-R models.

Quantitatively, the UNET-R model outperformed UNET-V, achieving a normalized cross-correlation of approximately 0.6 with the ground truth. In standard test cases, errors in density field metrics and relative velocity statistics were consistently below 10%.

Statistical Properties and Generalization

Beyond voxel-by-voxel accuracy, a critical finding was the model's ability to recover complex statistical metrics. The Radial Distribution Function (RDF), $g(r)$, which describes the probability of finding particle pairs at a given distance, was captured with high fidelity despite the model not being explicitly trained on this metric. This suggests that the spatial training objective effectively encodes the underlying statistical physics of the system.

Radial distribution functions g(r) and their ratios across different physical regimes, highlighting the model's statistical accuracy and generalization limits.

However, the model's generalization capabilities have clear boundaries. While robust to changes in the driving scales of turbulence, the model struggled when the energy injection rate was altered. Most significantly, it failed to generalize when the particle stopping time ($\tau_s$)—a fundamental parameter representing the friction between gas and dust—was changed. This indicates that while the U-Net learns a specific physical mapping, it does not yet inherently "understand" how changing physical constants alters that mapping.

Conclusion and Future Directions

This study demonstrates that 3D U-Nets can serve as powerful surrogate models for predicting particle clustering in turbulence, providing accurate spatial and statistical results with a fraction of the computational cost of traditional simulations.

Future work aims to address current limitations, such as the loss of sub-grid information inherent in grid-based mapping. The development of "Conditional U-Nets," which incorporate physical parameters like $\tau_s$ as additional inputs, could allow a single model to generalize across different dust sizes. Furthermore, extending this approach to "two-way coupling"—where dense dust regions exert a back-reaction on the gas—remains a vital frontier for applying deep learning to complex astrophysical fluid dynamics.

Summary updated

BibTeX
@article{chan2023particle,
  author = {Yan-Mong Chan and Natascha Manger and Yin Li and Chao-Chin Yang and Zhaohuan Zhu and Philip J. Armitage and Shirley Ho},
  title = {{Particle clustering in turbulence: Prediction of spatial and statistical properties with deep learning}},
  journal = {The Astrophysical Journal},
  year = {2023},
  doi = {10.3847/1538-4357/ad088c},
}
Figure